Search

MLOps es una metodología para la ingeniería ML (Machine Learning) que unifica el desarrollo del sistema ML (el elemento ML) con las operaciones del sistema ML (el elemento Ops). Aboga por formalizar y (cuando sea beneficioso) automatizar los pasos críticos de la construcción del sistema ML. MLOps proporciona un conjunto de procesos estandarizados y capacidades tecnológicas para construir, implementar y poner en funcionamiento sistemas ML de forma rápida y confiable. MLOps admite el desarrollo y la implementación de ML de la misma manera que DevOps y DataOps admiten la ingeniería de aplicaciones y la ingeniería de datos (análisis). La diferencia es que cuando implementa un servicio web, le importa la resiliencia, las consultas por segundo, el equilibrio de carga, etc. Cuando implementa un modelo ML, también debe preocuparse por los cambios en los datos, los cambios en el modelo y los usuarios que intentan trabajar con el sistema. MLOps es una metodología absolutamente novedosa (2020) y de escasa implementación en el tejido empresarial, en parte por lo reciente de sus conceptos, y en parte por la falta de un ecosistema de aplicaciones y métodos perfectamente establecidos que resulten de confianza para las compañías. Por el contrario, el ecosistema de herramientas y metodologías existente es variado y disperso, sin un líder establecido. Todo ello justifica el interés de realizar en este proyecto un análisis de las herramientas disponibles, metodologías y flujos de trabajo, para establecer las opciones más adecuadas en función de diferentes usos y organizaciones (grupo de investigación, PIME, gran empresa, etc).

MLOps is becoming a widespread practice in modern machine learning and data science. The word ”MLOps” combines machine learning technology and business operation process. Many business companies are applying machine learning techniques to improve their business targets and increase their profits. However, machine learning tasks involve complex applications and many technical stakeholders to perform a single system. It is not easy for non-technical operation staff to understand and proceed with the data science and machine learning processes without the aid of technical stakeholders. The communication effort between different stakeholders costs highly since the entire machine learning and operation processes are redundant and complicated. Therefore, MLOps provides insight for businesses to simplify the system workflow. MLOps pipeline automation simplified the processes in data, model, and production perspectives. This project researches several kinds of literature to identify the process of MLOps, including the phases and processes of MLOps, the necessary components for pipeline automation, and the concept of continuous integration and continuous delivery with the aid of MLOps. Also, two methods were selected to introduce a general MLOps pipeline design and implementation and the software tools and technologies selected in each MLOps process. Finally, a comparison section was provided to evaluate different software tools and technologies in five standard criteria., MLOps är på väg att bli en utbredd praxis inom modern maskininlärning och datavetenskap. Ordet ”MLOps” kombinerar maskininlärningsteknik (machine learning technology) och affärsprocess (business operation process). Många företag tillämpar maskininlärningstekniker för att förbättra sina affärsmål och öka sin vinst. Maskininlärningsuppgifter involverar dock komplexa applikationer och många tekniska intressenter. Det är inte lätt för icke-teknisk driftpersonal att förstå och gå vidare med datavetenskaps- och maskininlärningsprocesser utan hjälp av tekniska intressenter. Kommunikationsarbetet mellan olika intressenter är dyrt eftersom maskininlärnings- och driftprocesserna är repetitiva och komplicerade. MLOps förser företag med insikter som förenklar systemets arbetsflöde. Pipelineautomation med MLOps har förenklat processerna ur data-, modell- och produktionsperspektiv. Detta projekt undersöker flera typer av litteratur för att identifiera processen som MLOps tillämpar, inklusive faserna och processerna för MLOps, de nödvändiga komponenterna för pipeline automation samt konceptet kontinuerlig integration och kontinuerlig leverans med hjälp av MLOps. Dessutom valdes två metoder ut för att introducera en allmän MLOps-pipelinedesign och - implementation samt de mjukvaruverktyg och teknologier som valdes i varje MLOps-process. Slutligen tillhandahölls ett jämförelseavsnitt för att utvärdera olika mjukvaruverktyg och teknologier utifrån fem standardkriterier.

The burgeoning field of Artificial Intelligence (AI) is revolutionizing numerous industries, with machine learning (ML) models forming the core of many intelligent systems. However, transitioning effective ML models from development to production environments poses significant challenges. This research investigates the integration of Machine Learning Operations (MLOps) and DevOps principles, leveraging Artificial Intelligence (AI) to automate critical aspects of model deployment, version control, and lifecycle management. By streamlining the entire machine learning workflow, this approach aims to enhance the efficiency, reliability, and governance of AI-powered solutions. The paper commences with a comprehensive overview of MLOps and DevOps, highlighting their distinct yet complementary roles. MLOps encompasses a set of practices designed specifically for the unique challenges associated with the development, deployment, and management of ML models. These challenges include data versioning, model interpretability, performance monitoring, and drift detection. DevOps, on the other hand, focuses on fostering collaboration and communication between development and operations teams within the software development lifecycle. Its core principles of continuous integration/continuous delivery (CI/CD) facilitate rapid application delivery and infrastructure management. The paper then delves into the potential of AI for bridging the gap between MLOps and DevOps. AI techniques hold immense promise for automating various stages of the machine learning workflow. One crucial area of focus is automated model deployment. Traditionally, deploying ML models involves manual configuration and scripting, a time-consuming and error-prone process. AI-powered automation platforms can streamline this process by intelligently selecting target environments, provisioning resources, and configuring infrastructure based on model requirements. This not only reduces deployment time but also minimiz